ipex-llm 2.2.0b20250206__py3-none-win_amd64.whl → 2.2.0b20250208__py3-none-win_amd64.whl

ipex_llm/transformers/npu_pipeline_model/llama.py

@@ -18,108 +18,8 @@
 import torch
 import numpy as np
 import os
-from .common import update_names_of_IR_and_export_blob, LLMEmbedding, LowBitLLMLMHead, \
-    obtain_weight_from_single_layer
-from intel_npu_acceleration_library.backend.factory import NNFactory
-
-
-class Llama32Embedding(NNFactory):
-    def __init__(
-        self,
-        vocab_size,
-        embedding_dim,
-        embedding_weight,
-        padding_idx,
-        inv_freq,
-        attention_scaling,
-        dtype,  # fp16
-        device: str = "NPU",
-    ):
-        super().__init__(False, device)
-        self.vocab_size = vocab_size
-        self.embedding_dim = embedding_dim
-        self.padding_idx = padding_idx
-        self.attention_scaling = attention_scaling
-        self.dtype = dtype
-
-        # define input
-        weight = self.constant(embedding_weight)
-        input = self.parameter((1, 1), dtype=np.int32)
-        position_ids = self.parameter((1, 1), dtype=np.int64)
-        inv_freq = self.constant(inv_freq)
-
-        # embed_tokens module
-        if padding_idx == -1:
-            padding_idx += vocab_size
-
-        axis_node = self.constant(np.array([0], dtype=np.int64))
-        if padding_idx is not None:
-            masked_embeddings = np.ones(weight.shape, dtype=np.float16)
-            masked_embeddings[padding_idx, :] = 0.0  # mask
-
-            node_mask = self.constant(masked_embeddings)
-            node_masked_w = self.eltwise_mul(weight, node_mask)
-            res = self.gather(node_masked_w, input, axis_node, 0)
-        else:
-            res = self.gather(weight, input, axis_node, 0)
-
-        # rotary_emb module
-        inv_freq = self.reshape(inv_freq, (1, inv_freq.shape[0], 1))
-        position_ids = self.reshape(position_ids, (1, 1, 1))
-        freqs = self.eltwise_mul(self.convert_to_fp32(inv_freq),
-                                 self.convert_to_fp32(position_ids))
-        freqs = self.transpose(freqs, [0, 2, 1])
-        emb = self.concat(freqs, freqs, axis=2)
-        cos = self.cos(emb)
-        sin = self.sin(emb)
-        cos = cos * self.attention_scaling
-        sin = sin * self.attention_scaling
-
-        # define outputs
-        res = self.convert_to_fp16(res)
-        cos = self.convert_to_fp32(cos)
-        sin = self.convert_to_fp32(sin)
-
-        print("start compiling")
-        self.compile()
-
-
-class Llama32PostEmbedding(NNFactory):
-    def __init__(
-        self,
-        inv_freq,
-        attention_scaling,
-        input_len: int = 1,
-        device: str = "NPU",
-    ):
-        super().__init__(False, device)
-        self.attention_scaling = attention_scaling
-
-        # define input
-        position_ids = self.parameter((1, input_len), dtype=np.int64)
-        inv_freq = self.constant(inv_freq)
-
-        # rotary_emb module
-        inv_freq = self.reshape(inv_freq, (1, inv_freq.shape[0], 1))
-        position_ids = self.reshape(position_ids, (1, 1, input_len))
-        freqs = self.eltwise_mul(self.convert_to_fp32(inv_freq),
-                                 self.convert_to_fp32(position_ids))
-        freqs = self.transpose(freqs, [0, 2, 1])
-        emb = self.concat(freqs, freqs, axis=2)
-        cos = self.cos(emb)
-        sin = self.sin(emb)
-        cos = cos * self.attention_scaling
-        sin = sin * self.attention_scaling
-        if input_len > 1:
-            cos = self.unsqueeze(cos, [1])
-            sin = self.unsqueeze(sin, [1])
-
-        # define outputs
-        cos = self.convert_to_fp32(cos)
-        sin = self.convert_to_fp32(sin)
-
-        print("start compiling")
-        self.compile()
+from .common import update_names_of_IR_and_export_blob, LowBitLLMLMHead, \
+    obtain_weight_from_single_layer, obtain_embedding_from_model
 
 
 def convert_lm_head_and_embedding(model, n_splits_linear, temp_dir, weight_dir,
@@ -197,69 +97,17 @@ def convert_lm_head_and_embedding(model, n_splits_linear, temp_dir, weight_dir,
         bin_file = os.path.join(weight_dir, f"model_lm_head_input_{1+idx}.bin")
         weight.tofile(bin_file)
 
-    if hasattr(model.model.layers[0].self_attn.rotary_emb, "cos_cached"):
-        # llama-2-7B & llama-3-8B
-        embedding_layer = model.model.embed_tokens
-        new_embedding = LLMEmbedding(
-            vocab_size=model.config.vocab_size,
-            embedding_dim=model.config.hidden_size,
-            embedding_weight=embedding_layer.weight.to(torch.float16).detach().numpy(),
-            padding_idx=model.config.pad_token_id,
-            dtype=np.float16,
-        )
-        if convert_model:
-            bin_file = os.path.join(weight_dir, f"model_embedding_input_0.bin")
-            embedding_layer.weight.to(torch.float16).detach().numpy().tofile(bin_file)
-            first_blob_path = None
-        else:
-            first_blob_path = update_names_of_IR_and_export_blob(new_embedding, "embedding",
-                                                                 temp_dir, keep_ir=keep_ir,
-                                                                 compile_blob=compile_blob)
-            os.remove(os.path.join(temp_dir, "embedding.bin"))
-    else:
-        # llama-3.2-3B & llama-3.2-1B
-        embedding_layer = model.model.embed_tokens
-        new_embedding = Llama32Embedding(
-            vocab_size=model.config.vocab_size,
-            embedding_dim=model.config.hidden_size,
-            embedding_weight=embedding_layer.weight.to(torch.float16).detach().numpy(),
-            padding_idx=model.config.pad_token_id,
-            inv_freq=model.model.rotary_emb.inv_freq.to(torch.float16),
-            attention_scaling=model.model.rotary_emb.attention_scaling,
-            dtype=np.float16,
-        )
-        if convert_model:
-            bin_file = os.path.join(weight_dir, f"model_embedding_input_0.bin")
-            embedding_layer.weight.to(torch.float16).detach().numpy().tofile(bin_file)
-            first_blob_path = None
-            # save embedding post module
-            inv_freq = model.model.rotary_emb.inv_freq.to(torch.float16)
-            attention_scaling = model.model.rotary_emb.attention_scaling
-            embedding_post = Llama32PostEmbedding(inv_freq=inv_freq,
-                                                  attention_scaling=attention_scaling,
-                                                  input_len=1)
-            update_names_of_IR_and_export_blob(embedding_post, "embedding_post",
-                                               temp_dir, keep_ir=keep_ir, compile_blob=compile_blob)
-            embedding_post_prefill = Llama32PostEmbedding(inv_freq=inv_freq,
-                                                          attention_scaling=attention_scaling,
-                                                          input_len=max_prompt_len)
-            update_names_of_IR_and_export_blob(embedding_post_prefill,
-                                               "embedding_post_prefill",
-                                               temp_dir, keep_ir=keep_ir, compile_blob=compile_blob)
-            os.remove(os.path.join(temp_dir, "embedding_post.bin"))
-            os.remove(os.path.join(temp_dir, "embedding_post_prefill.bin"))
-        else:
-            first_blob_path = update_names_of_IR_and_export_blob(new_embedding, "embedding",
-                                                                 temp_dir, keep_ir=keep_ir,
-                                                                 compile_blob=compile_blob)
-            os.remove(os.path.join(temp_dir, "embedding.bin"))
+    first_blob_path = obtain_embedding_from_model(model, convert_model,
+                                                  temp_dir, weight_dir,
+                                                  max_prompt_len,
+                                                  keep_ir, compile_blob)
 
     return first_blob_path, last_blob_path
 
 
 def convert_llama_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
                         temp_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                        layernorm_const, mode="decode",
+                        const_parameter, mode="decode",
                         keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -297,14 +145,14 @@ def convert_llama_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     else:
         input_len = kv_len
         decoder_name = "decoder_layer_prefill"
-        layernorm_const = False
+        const_parameter = False
         keep_position_ids = False
         npu_dpu_groups = 6
 
     single_decoder = LowBitLlamaMultiDecoderlayer(
         [1, input_len, num_heads * head_dim],
-        input_layernorm_weights=[layer_norm_0] if layernorm_const else None,
-        post_attn_layernorm_weights=[layer_norm_1] if layernorm_const else None,
+        input_layernorm_weights=[layer_norm_0] if const_parameter else None,
+        post_attn_layernorm_weights=[layer_norm_1] if const_parameter else None,
         cached_cos=cached_cos,
         cached_sin=cached_sin,
         num_heads=num_heads,
@@ -334,7 +182,7 @@ def convert_llama_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     if mode == "decode":
         if hasattr(curr_layer.self_attn.rotary_emb, "cos_cached"):
             # llama-2-7B & llama-3-8B
-            if layernorm_const:
+            if const_parameter:
                 st_idx = 5
             else:
                 input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_3.bin")
@@ -344,7 +192,7 @@ def convert_llama_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
                 st_idx = 7
         else:
             # llama-3.2-3B & llama-3.2-1B
-            if layernorm_const:
+            if const_parameter:
                 st_idx = 6
             else:
                 input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_4.bin")
@@ -375,7 +223,7 @@ def convert_llama_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
 
 def convert_fused_llama_layer(model, fused_layers, n_splits_linear, n_splits_down_proj,
                               save_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                              layernorm_const, mode="decode",
+                              const_parameter, mode="decode",
                               keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -446,6 +294,10 @@ def convert_fused_llama_layer(model, fused_layers, n_splits_linear, n_splits_dow
         else:  # FP16 Linear
             np_dtype = np.float16
 
+        if not const_parameter:
+            input_layer_norm_weights = None
+            post_attn_layernorm_weights = None
+
         fused_decoder = LowBitLlamaMultiDecoderlayer(
             [1, 1, num_heads * head_dim],
             input_layernorm_weights=input_layer_norm_weights,

ipex_llm/transformers/npu_pipeline_model/minicpm.py

@@ -301,7 +301,7 @@ def convert_lm_head_and_embedding(model, n_splits_linear, temp_dir, weight_dir,
 
 def convert_minicpm_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
                           temp_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                          layernorm_const, mode="decode",
+                          const_parameter, mode="decode",
                           keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -333,12 +333,12 @@ def convert_minicpm_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     else:
         input_len = kv_len
         decoder_name = "decoder_layer_prefill"
-        layernorm_const = False
+        const_parameter = False
 
     single_decoder = LowBitMinicpmMultiDecoderlayer(
         [1, input_len, num_heads * head_dim],
-        input_layernorm_weights=[layer_norm_0] if layernorm_const else None,
-        post_attn_layernorm_weights=[layer_norm_1] if layernorm_const else None,
+        input_layernorm_weights=[layer_norm_0] if const_parameter else None,
+        post_attn_layernorm_weights=[layer_norm_1] if const_parameter else None,
         cached_cos=cached_cos,
         cached_sin=cached_sin,
         num_heads=num_heads,
@@ -364,7 +364,7 @@ def convert_minicpm_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     os.remove(os.path.join(temp_dir, decoder_name + ".bin"))
 
     if mode == "decode":
-        if layernorm_const:
+        if const_parameter:
             st_idx = 5
         else:
             input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_3.bin")
@@ -394,7 +394,7 @@ def convert_minicpm_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
 
 def convert_fused_minicpm_layer(model, fused_layers, n_splits_linear, n_splits_down_proj,
                                 save_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                                layernorm_const, mode="decode",
+                                const_parameter, mode="decode",
                                 keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -461,6 +461,10 @@ def convert_fused_minicpm_layer(model, fused_layers, n_splits_linear, n_splits_d
         else:  # FP16 Linear
             np_dtype = np.float16
 
+        if not const_parameter:
+            input_layer_norm_weights = None
+            post_attn_layernorm_weights = None
+
         fused_decoder = LowBitMinicpmMultiDecoderlayer(
             [1, 1, num_heads * head_dim],
             input_layernorm_weights=input_layer_norm_weights,

ipex_llm/transformers/npu_pipeline_model/qwen.py

@@ -18,13 +18,14 @@
 import torch
 import numpy as np
 import os
-from .common import update_names_of_IR_and_export_blob, LLMEmbedding, LowBitLLMLMHead, \
-    obtain_weight_from_single_layer, obtain_qkv_bias_from_single_layer
+from .common import update_names_of_IR_and_export_blob, LowBitLLMLMHead, \
+    obtain_weight_from_single_layer, obtain_qkv_bias_from_single_layer, \
+    obtain_embedding_from_model
 from ipex_llm.transformers.npu_models.lm_head import SlicedLMHead
 
 
 def convert_lm_head_and_embedding(model, temp_dir, weight_dir,
-                                  convert_model=False, group_size=0,
+                                  convert_model=False, group_size=0, max_prompt_len=1,
                                   keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     head_dim = model.model.layers[0].self_attn.head_dim
@@ -107,30 +108,16 @@ def convert_lm_head_and_embedding(model, temp_dir, weight_dir,
         bin_file = os.path.join(weight_dir, f"model_lm_head_input_{1+idx}.bin")
         weight.tofile(bin_file)
 
-    embedding_layer = model.model.embed_tokens
-    new_embedding = LLMEmbedding(
-        vocab_size=model.config.vocab_size,
-        embedding_dim=model.config.hidden_size,
-        embedding_weight=embedding_layer.weight.to(torch.float16).detach().numpy(),
-        padding_idx=model.config.pad_token_id,
-        dtype=np.float16,
-        input_length=1,
-    )
-    if convert_model:
-        bin_file = os.path.join(weight_dir, f"model_embedding_input_0.bin")
-        embedding_layer.weight.to(torch.float16).detach().numpy().tofile(bin_file)
-        first_blob_path = True
-    else:
-        first_blob_path = update_names_of_IR_and_export_blob(new_embedding, f"embedding",
-                                                             temp_dir, keep_ir=keep_ir,
-                                                             compile_blob=compile_blob)
-        os.remove(os.path.join(temp_dir, "embedding.bin"))
+    first_blob_path = obtain_embedding_from_model(model, convert_model,
+                                                  temp_dir, weight_dir,
+                                                  max_prompt_len,
+                                                  keep_ir, compile_blob)
     return first_blob_path, last_blob_path
 
 
 def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
                        temp_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                       layernorm_const, mode="decode",
+                       const_parameter, mode="decode",
                        keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -145,8 +132,13 @@ def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     mlp_layer = curr_layer.mlp
     weights = obtain_weight_from_single_layer(attn_layer, mlp_layer)
     q_bias, k_bias, v_bias = obtain_qkv_bias_from_single_layer(attn_layer)
-    cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
-    cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
+    if hasattr(curr_layer.self_attn.rotary_emb, "cos_cached"):
+        cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
+        cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
+    else:
+        # transformers >= 4.45.0
+        cached_cos = None
+        cached_sin = None
     layer_norm_0 = curr_layer.input_layernorm.weight.to(torch.float16)
     layer_norm_1 = curr_layer.post_attention_layernorm.weight.to(torch.float16)
 
@@ -158,10 +150,12 @@ def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
     if mode == "decode":
         input_len = 1
         decoder_name = f"decoder_layer_{layer_idx}"
+        keep_position_ids = True
         npu_dpu_groups = None
     else:
         input_len = kv_len
         decoder_name = "decoder_layer_prefill"
+        keep_position_ids = False
         npu_dpu_groups = 6
 
     single_decoder = LowBitQwenMultiDecoderlayer(
@@ -185,6 +179,8 @@ def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
         n_splits_linear=n_splits_linear,
         n_splits_down_proj=n_splits_down_proj,
         group_size=group_size,
+        cos_len=input_len,
+        keep_position_ids=keep_position_ids,
         asym=asym
     )
     rest_blob_path = update_names_of_IR_and_export_blob(single_decoder,
@@ -196,14 +192,25 @@ def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
 
     # 0, 1, 2 are input_embed/attention_mask/position_id
     if mode == "decode":
-        if layernorm_const:
-            st_idx = 3
+        if hasattr(curr_layer.self_attn.rotary_emb, "cos_cached"):
+            if const_parameter:
+                st_idx = 3
+            else:
+                input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_3.bin")
+                post_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_4.bin")
+                layer_norm_0.data.numpy().tofile(input_lm_bin_file)
+                layer_norm_1.data.numpy().tofile(post_lm_bin_file)
+                st_idx = 5
         else:
-            input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_3.bin")
-            post_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_4.bin")
-            layer_norm_0.data.numpy().tofile(input_lm_bin_file)
-            layer_norm_1.data.numpy().tofile(post_lm_bin_file)
-            st_idx = 5
+            # transformers >= 4.45.0
+            if const_parameter:
+                st_idx = 4
+            else:
+                input_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_4.bin")
+                post_lm_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_5.bin")
+                layer_norm_0.data.numpy().tofile(input_lm_bin_file)
+                layer_norm_1.data.numpy().tofile(post_lm_bin_file)
+                st_idx = 6
         q_bias_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_{st_idx}.bin")
         k_bias_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_{st_idx+1}.bin")
         v_bias_bin_file = os.path.join(weight_dir, f"model_{layer_idx}_input_{st_idx+2}.bin")
@@ -234,7 +241,7 @@ def convert_qwen_layer(model, layer_idx, n_splits_linear, n_splits_down_proj,
 
 def convert_fused_qwen_layer(model, fused_layers, n_splits_linear, n_splits_down_proj,
                              save_dir, weight_dir, transpose_value_cache, kv_len, group_size,
-                             layernorm_const, mode="decode",
+                             const_parameter, mode="decode",
                              keep_ir=False, compile_blob=True):
     num_heads = model.model.layers[0].self_attn.num_heads
     num_key_value_heads = model.model.layers[0].self_attn.num_key_value_heads
@@ -261,8 +268,13 @@ def convert_fused_qwen_layer(model, fused_layers, n_splits_linear, n_splits_down
             attn_layer = curr_layer.self_attn
             mlp_layer = curr_layer.mlp
             weights = obtain_weight_from_single_layer(attn_layer, mlp_layer)
-            cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
-            cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
+            if hasattr(curr_layer.self_attn.rotary_emb, "cos_cached"):
+                cached_cos = curr_layer.self_attn.rotary_emb.cos_cached.to(torch.float16)
+                cached_sin = curr_layer.self_attn.rotary_emb.sin_cached.to(torch.float16)
+            else:
+                # transformers >= 4.45.0
+                cached_cos = None
+                cached_sin = None
             layer_norm_0 = curr_layer.input_layernorm.weight.to(torch.float16)
             layer_norm_1 = curr_layer.post_attention_layernorm.weight.to(torch.float16)
 
@@ -313,6 +325,13 @@ def convert_fused_qwen_layer(model, fused_layers, n_splits_linear, n_splits_down
         else:  # FP16 Linear
             np_dtype = np.float16
 
+        if not const_parameter:
+            input_layer_norm_weights = None
+            post_attn_layernorm_weights = None
+            q_biases = None
+            k_biases = None
+            v_biases = None
+
         fused_decoder = LowBitQwenMultiDecoderlayer(
             [1, 1, num_heads * head_dim],
             input_layernorm_weights=input_layer_norm_weights,

{ipex_llm-2.2.0b20250206.dist-info → ipex_llm-2.2.0b20250208.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250206
+Version: 2.2.0b20250208
 Summary: Large Language Model Develop Toolkit
 Home-page: https://github.com/intel-analytics/ipex-llm
 Author: BigDL Authors
@@ -27,19 +27,12 @@ Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'all'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'all'
 Provides-Extra: cpp
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250206 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250208 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
-Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250206 ; extra == 'cpp-arl'
-Requires-Dist: setuptools ; extra == 'cpp-arl'
-Requires-Dist: onednn-devel ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
-Requires-Dist: onednn ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
-Requires-Dist: dpcpp-cpp-rt ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
-Requires-Dist: mkl-dpcpp ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
-Requires-Dist: onednn-devel ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp'
-Requires-Dist: onednn ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp'
-Requires-Dist: dpcpp-cpp-rt ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp'
-Requires-Dist: mkl-dpcpp ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp'
+Requires-Dist: onednn-devel ==2025.0.1 ; (platform_system == "Windows") and extra == 'cpp'
+Requires-Dist: onednn ==2025.0.1 ; (platform_system == "Windows") and extra == 'cpp'
+Requires-Dist: dpcpp-cpp-rt ==2025.0.2 ; (platform_system == "Windows") and extra == 'cpp'
+Requires-Dist: mkl-dpcpp ==2025.0.1 ; (platform_system == "Windows") and extra == 'cpp'
 Provides-Extra: llama-index
 Requires-Dist: py-cpuinfo ; extra == 'llama-index'
 Requires-Dist: protobuf ; extra == 'llama-index'
@@ -67,7 +60,7 @@ Requires-Dist: transformers ==4.40.0 ; extra == 'npu'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'npu'
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'npu'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'npu'
-Requires-Dist: bigdl-core-npu ==2.6.0b20250206 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250208 ; (platform_system == "Windows") and extra == 'npu'
 Provides-Extra: serving
 Requires-Dist: py-cpuinfo ; extra == 'serving'
 Requires-Dist: fschat[model_worker,webui] ==0.2.36 ; extra == 'serving'
@@ -87,9 +80,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250206 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250206 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250206 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250208 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250208 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250208 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -104,9 +97,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu-2-1'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250206 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250206 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250206 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250208 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250208 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250208 ; extra == 'xpu-2-1'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-1'
 Requires-Dist: dpcpp-cpp-rt ==2024.0.2 ; (platform_system == "Windows") and extra == 'xpu-2-1'
 Requires-Dist: mkl-dpcpp ==2024.0.0 ; (platform_system == "Windows") and extra == 'xpu-2-1'
@@ -124,7 +117,7 @@ Requires-Dist: setuptools ; extra == 'xpu-2-6'
 Requires-Dist: torch ==2.6.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchvision ==0.21.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchaudio ==2.6.0+xpu ; extra == 'xpu-2-6'
-Requires-Dist: bigdl-core-xe-all ==2.6.0b20250206 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250208 ; extra == 'xpu-2-6'
 Requires-Dist: onednn-devel ==2025.0.1 ; extra == 'xpu-2-6'
 Requires-Dist: onednn ==2025.0.1 ; extra == 'xpu-2-6'
 Requires-Dist: dpcpp-cpp-rt ==2025.0.2 ; extra == 'xpu-2-6'
@@ -140,9 +133,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arc'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arc'
 Requires-Dist: tabulate ; extra == 'xpu-arc'
 Requires-Dist: setuptools ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250206 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250206 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250206 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250208 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250208 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250208 ; extra == 'xpu-arc'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arc'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
@@ -163,9 +156,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arl'
 Requires-Dist: tabulate ; extra == 'xpu-arl'
 Requires-Dist: setuptools ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250206 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250206 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250206 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250208 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250208 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250208 ; extra == 'xpu-arl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
@@ -186,9 +179,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-lnl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-lnl'
 Requires-Dist: tabulate ; extra == 'xpu-lnl'
 Requires-Dist: setuptools ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250206 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250206 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250206 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250208 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250208 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250208 ; extra == 'xpu-lnl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-lnl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'